April 17, 2017

Hi!

Daniel Chen

Computational. Network. Epidemiology.

About Me

Teach

Published Teaching

Current Work

Modeling Attitude Diffusion

Modeling

(Computational) (Infectious) Disease Modeling

Two main types of models:

  1. Compartmental/Mathematical/System Dynamics/ Ordinary Differential Equation (ODE) Models
  2. Agent-Based Models (ABM)

Compartmental Models

SIR Models

  • Susceptible
  • Infectious
  • Recovered

\[ \begin{align*} \frac{dS}{dt} &= -b S(t) I(t)\\ \frac{dI}{dt} &= b S(t) I(t) - kI(t)\\ \frac{dR}{dt} &= k I(t) \end{align*} \]

ISEE Systems - Stella

Compartmental Models

  • SIS
  • SEIR
  • MSIR
  • SI(CR)
  • Vaccine

Pros

  • Deterministic
  • Overall System Dynamics
  • Homogeneous (random mixing)
  • Simple, easy to set up

Cons

  • Not stochastic
  • No individual level behaviors
  • No complex interactions

Agent-Based Models

  • The model is composed of individual 'agents'
  • Each agent has a set of rules

  • The agents repeat these rules (ticks/cycles)

  • Observe complex system dynamics from the bottom-up through emergence

Agent-Based Models

Segregation

Wolf Sheep Predation

Agent-Based Models

Virus on a Network

Agent-Based Models

Pros

  • Stochastic
  • Consolidate knowledge for agent rules/behaviors
  • Heterogeneous

Cons

  • Needs a lot of data and time to set up
  • Harder to get general system dynamics
  • Needs a lot of model runs
  • Resource-intensive

CMs to ABMs

Pablo Picasso, "Bull", Plates 1-11 (Lithograph)

(Duncan) Watts Model

Watts 2002: A simple model of global cascades on random networks

  • Binary Decisions with Externalities (general contagion model)
    • fads, riots, crime, competing technologies, spread of innovation, conventions, and cooperation
  • Probability of a global cascade from a single node
  • Local dependencies, fractional threshold, and heterogeneity

Definitions

  • Cascades: event of any size triggered by an initial seed
  • Global cascades: a cascade that occupies a finite fraction of an infinite network. A sufficiently large cascade. More than a fixed fraction of a large, but finite network.
  • Local dependencies: agents will incorporate information about its neighbors
  • Fractional threshold: agents themselves a threshold that determine how it incorporates information from its neighbors
  • Heterogeneity: every agent is different to varying degree from one another

Analogy from the paper

Diffution of Innovations:

  • Innovators ~ initial seed

  • Early adopters ~ vulnerable nodes

  • A cascade will occur if innovators are connected to many early adopters (connectivity)

  • More early adopters, higher chance of innovation, but they need to be connected (structure)

Simulation

  • \(n\) agents in a network start off with a state of \(0\)
  • Individual agents can only have a state that is either \(0\) or \(1\)
  • Each agent has \(k\) neighbors
  • An agent gets a new state of \(1\) if a fraction of its neighbors, \(\phi\), are also \(1\)
    • Otherwise an agent gets a new state of \(0\)
  • During each time step, the population evolves:
    • Update states in random, asynchronous order using the threshold rule
    • Once an agent has a state of \(1\), it will stay at \(1\) for the remainder of the simulation

\(\phi\) and \(k\) are 2 parameters we can change

Simulation Parameterization

  • \(\phi\) and \(k\) may be heterogeneous
    • To simplify the simulations, the paper has a homogeneous threshold, \(\phi\)
  • The network is a uniform random graph
  • A small seed
  • Any pair of nodes is connected with probability \(p = \frac{z}{n}\)
    • in a uniform random graph, $p_k = $ Poisson distribution
  • \(n = 10,000\)
  • 100 random runs of each simulation

Watts Threshold

Replication Study

Replication Study

Expand the Watts Model

The Watts model can be used to model any binary outcome

From a public health and epidemiology perspective, this outcome can be a particular behavior or action.

  • I ate a cookie. Yes/No

However, our decision making process is not that simple.

Theory of Reasoned Action (TRA)

TRA Health behavior model

  • Martin Fishbein and Icek Ajzen 1967
  • Behavior is determined by an individual's intention
  • Intention comes from an individual's attitudes and social context
  • Attitudes and social context originates from a set of beliefs

beliefs \(\rightarrow\) (attitudes & social context) \(\rightarrow\) intention \(\rightarrow\) behavior

Neural Networks

  • Feed-forward
  • Recurrent
  • Convolutional
  • Psychological plausible decisions
    • social processes
    • experience/ memory
    • influences/ dynamic

Orr 2014 Figure 1

TRA as Parallel Constraint Satisfaction

Orr 2013 Figure 1

Orr 2013 Figure 2

Results

  • parameter sweep on NN connections
  • no learning during simulation
  • 7750 Simulations
  • associativity
    • clustering of intention at the end of the simulation

Future

  • Election 'Fake News'
  • NYT: Shooting Scares Show a Nation Quick to Fear the Worst

Questions: - How to summarize the NN Simulation data - Signal Decomposition/analysis - Projections (t-sne)

(Relevant?) Past Work

Wicer EntiCE3

  • Adriana Arcia, Mark Velez, Suzanne Bakken
  • Niurka Suero-Tejeda, Amenda Almonte
  • Programatically generate infographics
  • Improve health literacy

Main Lesson: Running software for non-technical people (shiny?)

Mass Casualty

  • 2014: Mass fatality preparedness among medical examiners/coroners in the United States: a cross-sectional study
    • Robyn RM Gershon, Mark G Orr, Qi Zhi, Jacqueline A Merrill, Daniel Y Chen, Halley EM Riley, and Martin F Sherman
  • 2015: Are We Ready for Mass Fatality Incidents? Preparedness of the US Mass Fatality Infrastructure
    • Jacqueline A. Merrill, Mark Orr, Daniel Y. Chen, Qi Zhi, and Robyn R. Gershon

Main Findings: "The sectors in the US mass fatality infrastructure report suboptimal capability to respond. National leadership is needed…"

Lessons Learned: Data quality governs analysis

Patterns of Care: Congestive Heart Failure

  • "Transition Networks in a Cohort of Patients with Congestive Heart Failure"
  • Jacqueline A. Merrill
  • Kathleen M. Carley

  • Barbara Sheehan

  • Looked for patients who had no re-admissions, re-admissions, and multiple re-admissions

Lessons Learned: EHR Data is really messy

Work

OpenX

Open Data Zika Dashboard

Open Data Brazil

Scientific Computing

Social and Decision Analytics Lab

  • Data Science Deployment Infrastructure
    • VirtualBox VMs
    • LXC Containers
    • Docker Containers

Thanks!